One-bit Supervision for Image Classification

Authors: Hengtong Hu, Lingxi Xie, Zewei Du, Richang Hong, Qi Tian

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate our setting and approach on three image classification benchmarks, namely, CIFAR100, Mini-Image Net and Image Net. ... Results demonstrate the superiority of one-bit supervision, and, with diagnostic experiments, we verify that the benefits come from a more efficient way of utilizing the information of incomplete supervision.
Researcher Affiliation Collaboration Hengtong Hu1,2, Lingxi Xie3, Zewei Du3 Richang Hong1,2 , Qi Tian3 1Key Laboratory of Knowledge Engineering with Big Data, Hefei University of Technology, 2School of Computer Science and Information Engineering, Hefei University of Technology, 3Huawei Inc.
Pseudocode No The paper describes the training procedure verbally and with a diagram (Figure 1), but does not provide structured pseudocode or an algorithm block.
Open Source Code No The paper does not provide an unambiguous statement about releasing source code for the described methodology or a direct link to a code repository.
Open Datasets Yes We conduct experiments on three popular image classification benchmarks, namely, CIFAR100, Mini-Imagenet, and Imagenet. CIFAR100 [16] contains 50K training images and 10K testing images... For Mini-Image Net in which the image resolution is 84 84, we use the training/testing split created in [28]... For Image Net [3], we use the commonly used competition subset [30]...
Dataset Splits No The paper describes training and test sets but does not explicitly mention or detail a separate validation dataset split with specific percentages or counts for hyperparameter tuning or model selection.
Hardware Specification Yes Other hyper-parameters simply follow the original implementation, except that the batch size is adjusted to fit our hardware (e.g., eight NVIDIA Tesla-V100 GPUs for Image Net experiments).
Software Dependencies No The paper refers to adopted algorithms and network architectures (e.g., Mean-Teacher, ResNet) and states that 'Other hyper-parameters simply follow the original implementation,' but it does not specify software dependencies with version numbers (e.g., PyTorch 1.9, TensorFlow 2.x).
Experiment Setup Yes We use a 26-layer deep residual network [10] with Shake-Shake regularization [7] for CIFAR100, and a 50-layer residual network for Mini-Image Net and Image Net. The number of training epochs is 180 for CIFAR100 and Mini-Image Net, and 60 for Image Net. The consistency parameter is 1,000 for CIFAR100, and 100 for Mini-Image Net and Image Net.